GRV-KBQA - A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure, Semantic Grounding and Structure-Aware Validation

Sep 19, 2026

Knowledge Base Question Answering (KBQA) retrieves answers from structured knowledge bases given a natural language question, yet existing retrieval, semantic parsing, and LLM-based approaches often generate non-executable queries and incur inefficient execution. GRV-KBQA introduces a three-stage framework that decouples logical structure generation from semantic grounding and adds structure-aware validation that enforces KB constraints. Experiments on WebQSP and CWQ demonstrate significant performance gains, and ablation confirms the effectiveness of the decoupled generation and validation mechanism.

LanguageEnglish

CategoryGeneral

FormatPDF

Pages15

Reading Time5 min

AccessFree to download

Introduction

Knowledge bases and KBQA task

Prior methods: IR-based and Semantic Parsing-based

Key challenges in logical form generation and KB alignment

Proposed approach: GRV-KBQA framework

GRV-KBQA: A Three-Stage Framework for Knowledge Base Question Answering with Decoupled Logical Structure, Semantic Grounding and Structure-Aware Validation Yuhang Tian1 * , Pan Yang1 * , Dandan Song1†, Zhijing Wu1 , Hao Wang1 1 School of Computer Science and Technology, Beijing Institute of Technology, China {tianyuhang,[sdd}](sdd}@bit.edu.cn)[@bit.edu.cn](sdd}@bit.edu.cn), [zaixiatongxin@gmail.com](zaixiatongxin@gmail.com) Abstract Knowledge Base Question Answering (KBQA) is a fundamental task that enables natural language interaction with structured knowledge bases (KBs) . Given a natural language question, KBQA aims to retrieve the answers from the KB. However, existing approaches, including retrieval-based, semantic parsing-based methods and large-language model-based methods often suffer from generating non-executable queries and inefficiencies in query execution. To address these challenges, we propose GRVKBQA, a three-stage framework that decouples logical structure generation from semantic grounding and incorporates structure-aware validation to enhance accuracy. Unlike previous methods, GRV-KBQA explicitly enforces KB constraints to improve alignment between generated logical forms and KB structures. Experimental results on WebQSP and CWQ show that GRV-KBQA significantly improves performance over existing approaches. The ablation study conducted confirms the effectiveness of the decoupled logical form generation and validation mechanism of our framework. 1 Introduction Knowledge bases (KBs) such as Freebase (Bollacker et al., 2008), Wikidata (Vrandei and Krötzsch, 2014), and DBpedia (Auer et al., 2007), which contain vast amounts of structured data in the form of triples, are widely used due to their structured nature and the accuracy of the information they represent. Knowledge Base Questioning Answering (KBQA) is a popular application of KB, aiming to retrieve accurate answers from the KB for a given question. Previous methods for Knowledge Base Question Answering (KBQA) can be broadly categorized into two main classes: Information Retrievalbased (IR-based) (Miller et al., 2016 ; Sun et al., *Equal contribution. †Corresponding author. Figure 1: Comparison of GRV-KBQA with previous methods. 2019 ; He et al., 2021a ; Zhang et al., 2022) methods and Semantic Parsing-based (SP-based) methods. IR-based methods primarily focus on retrieving relevant subgraphs from the KB and performing reasoning over these subgraphs to derive the answer. Besides, SP-based methods (Das et al., 2021 ; Lan and Jiang, 2020 ; Ye et al., 2022 ; Shu et al., 2022) aim to transform natural language questions into executable logical forms, such as SPARQL queries, which are then executed directly on the KB to obtain the answer. Currently, as the large language models (LLMs) have exceptional generative capabilities and powerful learning abilities, 2618 Findings of the Association for Computational Linguistics: EMNLP 2025 , pages 2618–2632 November 4-9, 2025 ©2025 Association for Computational Linguistics several research efforts have proposed sequence-tosequence (seq-to-seq) approaches based on LLMs, which directly generate logical forms, align them to the KB, and then execute them on the KB to retrieve answers. KB-Coder (Nie et al., 2024) and ARG-KBQA (Tian et al., 2024) have leveraged LLMs for few-shot in-context learning, where a small number of examples guides the model to generate logical forms. To fully leverage the capabilities of LLMs while avoiding the use of black-box models, ChatKBQA (Luo et al., 2024) fine-tunes open-source models to enable them to generate logical forms based on the given questions. However, existing methods still face two key issues: 1. Challenges in Generating Correct Logical Forms: They suffer from generating nonexecutable logical forms in many cases. For example, the basic logical structure of generated logical form is wrong, as shown in Figure 1. This limitation primarily arises f

FAQs

What problem does GRV-KBQA target in knowledge base question answering?

It targets two issues: generating incorrect, non-executable logical forms and inefficient alignment when executing those forms on the knowledge base.

How does GRV-KBQA improve logical form generation?

It uses decoupled logical structure generation with a two-phase compositional approach based on a subquery chain template, which reduces errors from one-step joint generation.

How does structure-aware validation help during query execution?

It enforces consistency with knowledge base structure, improving alignment between generated logical forms and KB constraints to avoid invalid queries.

Sep 19, 2026

Upload by Lucas Vance